Me-ow! Russian tech titan Yandex open-sources ML library CatBoost

Russia's tech behemoth Yandex has open-sourced its first machine learning library, CatBoost. Many big-name tech companies including Google, Facebook, Microsoft and even Sony already offer machine learning frameworks. These tend to focus on neural networks, computer systems modelled on the human brain that can be trained to …

[zarvus]

A+ for headline

[John Smith 19]

Obvious question with gradient optimization questions.

How does it cope with local, but not global, minima?

[Ben1892]

Re: Obvious question with gradient optimization questions.

I was thinking exactly the same !

(Actually I wasn't but I have recently decided to get myself up to speed with machine learning theory and you've given me something to read up about - cheers )

[John Smith 19]

Re: Obvious question with gradient optimization questions.

I think of it as a path up and down a series of hills.

Across the whole landscape, where is the real absolute minimum or maximum? On a screen it's more or less obvious, but a) Computers are blind and b) the graph may be an n-dimensional surface, which you can't show on a screen anyway.

So you use first and second derivatives to find zeros and inflexion points, but of course you probably don't have the equations for the data (which would let you find the real points analytically) so you have to do it numerically with the associated (potential) s**tstorm of numerical instability.

Obviously software packages should handle such problems (or let you know they are happening so you can handle them) but that's about the limit of my knowledge of the subject.